Method: compared GitHub Trending daily snapshots (Rain1er/trending, Sep 27 23:32 UTC; agents-radar, Sep 28 01:06 UTC) with the previous day’s snapshot (Sep 26 22:54 UTC). Excluded: paperclipai/paperclip (89k total), rohitg00/ai-engineering-from-scratch (59k), debpalash/VoiceStudio (+3,060 today but 39.9k total, with a GitHub Trending #1 on Sep 15 and a #1 weekly ranking in Week 36, so it is keeping its rank rather than newly rising), NVIDIA/Model-Optimizer (dropped off today’s list) and ultralytics (62k).
vectorize-io/hindsight — +4,463 ⭐ today (2 days in a row)
URL: https://github.com/vectorize-io/hindsight
Total: ~35.9k–37.4k ⭐ (Sep 27 +2,152 → Sep 28 +4,463–4,520, about 2.1x) · Python · MIT
Hindsight is an agent memory system that makes agents learn over time instead of just replaying conversation history. Its three core operations:
- Retain: uses an LLM to extract and store facts, entities, relationships and temporal information
- Recall: runs semantic search, keyword matching, entity/temporal/causal graph traversal and time filtering in parallel to retrieve memories
- Reflect: forms new connections among stored memories to build a synthesized understanding
Memories are split into world facts, experiences, observations (deduplicated beliefs strengthened by evidence) and mental models, closer to human memory than a plain vector database. It supports 25+ LLM providers, 60+ integrations including LangChain, CrewAI, Claude Code and Cursor, and an MCP server. After accelerating 30% yesterday, its daily gain doubled today, the biggest of any trending repo. With memory layers like mem0 and cognee also drawing attention, Hindsight is riding the view that keeping context across long-running agent sessions is the real bottleneck for production agents. Its total is nearing 40k, so it will likely drop out under the 40k rule from the next edition.
Practical use: Attach per-customer long-term memory to support or sales-assist agents so past inquiries, preferences and decisions carry over without re-explaining.
Tags: #AgentMemory #LongTermMemory #KnowledgeGraph #LongMemEval #ContextEngineering
dream-num/univer — +920 ⭐ today
URL: https://github.com/dream-num/univer
Total: ~20.0k–20.3k ⭐ (Sep 27 +845 → Sep 28 +895–920, rebounding) · TypeScript · Apache-2.0
Univer is an open-source office SDK that handles spreadsheets, documents, slides, databases (Bases) and canvas in a single runtime, recently repositioned as “the office harness for AI agents.” Every capability ships as a plugin, and the same Facade API runs in the browser (React 18+, canvas rendering) and headless in Node.js, backed by its own formula engine and a rendering engine for large documents. Ways agents can work with it:
- Facade API for programmatic editing, content inspection and rendered-screenshot verification
- Worktree-style isolated drafts where agents work and humans review
- MCP integration for natural-language control, and univer-cli for command-line agent workflows
- Agent skill documentation (@univerjs-sdk-skills)
After gaining more than 15,000 stars in a single day on September 23 and then slowing, it ticked back up today — a sign of continuing demand for “agentic office” tools that let agents edit spreadsheets and documents directly.
Practical use: Use it as the backend for internal automation that fills Excel or document templates such as RFPs and quotes, with agents editing while preserving formulas and formatting and verifying results via screenshots.
Tags: #AgenticOffice #SpreadsheetEngine #DocumentAutomation #HeadlessOffice #MCP
mvschwarz/openrig — +114 ⭐ today (new)
URL: https://github.com/mvschwarz/openrig
Total: 712 → ~990–1.1k ⭐ (up roughly 40–50% in a day) · TypeScript · Apache-2.0
OpenRig is a multi-agent harness that runs Claude Code and Codex together as one system. Instead of juggling terminal sessions, you declare a team in YAML (RigSpec) and boot the whole team with a single rig up. It runs on tmux as a local daemon, CLI, TUI and MCP server, storing state in SQLite. Key concepts:
- Seats: stable roles with persistent addresses such as dev-owner@first-project
- Pods: groups of seats sharing guidance and context
- Messaging between agents, plus discovering and adopting existing Claude or Codex sessions
- Topology snapshots and restore, and portable bundles with vendored agent blueprints
- MCP tools that let agents launch teams and check status themselves
Starter templates include product-team, conveyor (a handoff pipeline) and research/implementation pairs. A small repo of about 1k stars making its first appearance on daily trending, it is this edition’s only pure newcomer and reflects the push to run several coding agents like an organization.
Practical use: Reproduce and share an internal dev-pipeline PoC that splits design, implementation and review between Claude Code and Codex from a single YAML file.
Tags: #MultiAgentOrchestration #ClaudeCode #Codex #Tmux #AgentHarness #MCP
Leave a comment